-
Notifications
You must be signed in to change notification settings - Fork 18
/
Copy pathdataset_aqi36.py
193 lines (169 loc) · 8.69 KB
/
dataset_aqi36.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
import pickle
from torch.utils.data import DataLoader, Dataset
import pandas as pd
import numpy as np
import torch
import torchcde
from utils import get_randmask, get_hist_mask
class AQI36_Dataset(Dataset):
def __init__(self, eval_length=36, target_dim=36, mode="train", val_len=0.1, is_interpolate=False,
target_strategy='hybrid', mask_sensor=None, missing_ratio=None):
self.eval_length = eval_length
self.target_dim = target_dim
self.is_interpolate = is_interpolate
self.target_strategy = target_strategy
self.mode = mode
self.missing_ratio = missing_ratio
self.mask_sensor = mask_sensor
path = "./data/pm25/pm25_meanstd.pk"
with open(path, "rb") as f:
self.train_mean, self.train_std = pickle.load(f)
if mode == "train":
month_list = [1, 2, 4, 5, 7, 8, 10, 11]
# 1st,4th,7th,10th months are excluded from histmask (since the months are used for creating missing patterns in test dataset)
flag_for_histmask = [0, 1, 0, 1, 0, 1, 0, 1]
elif mode == "valid":
month_list = [2, 5, 8, 11]
elif mode == "test":
month_list = [3, 6, 9, 12]
self.month_list = month_list
# create data for batch
self.observed_data = [] # values (separated into each month)
self.observed_mask = [] # masks (separated into each month)
self.gt_mask = [] # ground-truth masks (separated into each month)
self.index_month = [] # indicate month
self.position_in_month = [] # indicate the start position in month (length is the same as index_month)
self.valid_for_histmask = [] # whether the sample is used for histmask
self.use_index = [] # to separate train/valid/test
self.cut_length = [] # excluded from evaluation targets
df = pd.read_csv(
"./data/pm25/SampleData/pm25_ground.txt",
index_col="datetime",
parse_dates=True,
)
df_gt = pd.read_csv(
"./data/pm25/SampleData/pm25_missing.txt",
index_col="datetime",
parse_dates=True,
)
for i in range(len(month_list)):
current_df = df[df.index.month == month_list[i]]
current_df_gt = df_gt[df_gt.index.month == month_list[i]]
if mode == 'train' and month_list[i] in [2, 5, 8, 11]:
cut_len = int(val_len * len(current_df))
current_df = current_df[:-cut_len]
current_df_gt = current_df_gt[:-cut_len]
if mode == 'valid':
cut_len = int(val_len * len(current_df))
current_df = current_df[-cut_len:]
current_df_gt = current_df_gt[-cut_len:]
current_length = len(current_df) - eval_length + 1
last_index = len(self.index_month)
self.index_month += np.array([i] * current_length).tolist()
self.position_in_month += np.arange(current_length).tolist()
if mode == "train":
self.valid_for_histmask += np.array(
[flag_for_histmask[i]] * current_length
).tolist()
# mask values for observed indices are 1
c_mask = 1 - current_df.isnull().values
c_gt_mask = 1 - current_df_gt.isnull().values
if len(self.mask_sensor) > 0:
for sensor in self.mask_sensor:
c_gt_mask[:, sensor] = 0
if self.mode == 'train':
for sensor in self.mask_sensor:
c_mask[:, sensor] = 0
c_data = (
(current_df.fillna(0).values - self.train_mean) / self.train_std
) * c_mask
self.observed_mask.append(c_mask)
self.gt_mask.append(c_gt_mask)
self.observed_data.append(c_data)
if mode == "test":
n_sample = len(current_df) // eval_length
# interval size is eval_length (missing values are imputed only once)
c_index = np.arange(
last_index, last_index + eval_length * n_sample, eval_length
)
self.use_index += c_index.tolist()
self.cut_length += [0] * len(c_index)
if len(current_df) % eval_length != 0: # avoid double-count for the last time-series
self.use_index += [len(self.index_month) - 1]
self.cut_length += [eval_length - len(current_df) % eval_length]
if mode != "test":
self.use_index = np.arange(len(self.index_month))
self.cut_length = [0] * len(self.use_index)
# masks for 1st,4th,7th,10th months are used for creating missing patterns in test data,
# so these months are excluded from histmask to avoid leakage
if mode == "train":
ind = -1
self.index_month_histmask = []
self.position_in_month_histmask = []
for i in range(len(self.index_month)):
while True:
ind += 1
if ind == len(self.index_month):
ind = 0
if self.valid_for_histmask[ind] == 1:
self.index_month_histmask.append(self.index_month[ind])
self.position_in_month_histmask.append(
self.position_in_month[ind]
)
break
else: # dummy (histmask is only used for training)
self.index_month_histmask = self.index_month
self.position_in_month_histmask = self.position_in_month
def __getitem__(self, org_index):
index = self.use_index[org_index]
c_month = self.index_month[index]
c_index = self.position_in_month[index]
index2 = np.random.randint(0, len(self.use_index))
hist_month = self.index_month_histmask[index2]
hist_index = self.position_in_month_histmask[index2]
ob_data = self.observed_data[c_month][c_index:c_index + self.eval_length]
ob_mask = self.observed_mask[c_month][c_index:c_index + self.eval_length]
ob_mask_t = torch.tensor(ob_mask).float()
gt_mask = self.gt_mask[c_month][c_index:c_index + self.eval_length]
for_pattern_mask = self.observed_mask[hist_month][hist_index:hist_index + self.eval_length]
if self.mode != 'train':
cond_mask = torch.tensor(gt_mask).to(torch.float32)
else:
if self.target_strategy != 'random':
cond_mask = get_hist_mask(ob_mask_t, for_pattern_mask=for_pattern_mask)
else:
cond_mask = get_randmask(ob_mask_t)
s = {
"observed_data": ob_data,
"observed_mask": ob_mask,
"gt_mask": gt_mask,
"hist_mask": for_pattern_mask,
"timepoints": np.arange(self.eval_length),
"cut_length": self.cut_length[org_index],
"cond_mask": cond_mask.numpy()
}
if self.is_interpolate:
tmp_data = torch.tensor(ob_data).to(torch.float64)
itp_data = torch.where(cond_mask == 0, float('nan'), tmp_data).to(torch.float32)
itp_data = torchcde.linear_interpolation_coeffs(
itp_data.permute(1, 0).unsqueeze(-1)).squeeze(-1).permute(1, 0)
s["coeffs"] = itp_data.numpy()
return s
def __len__(self):
return len(self.use_index)
def get_dataloader(batch_size, device, val_len=0.1, is_interpolate=False, num_workers=4, target_strategy='hybrid', mask_sensor=None):
dataset = AQI36_Dataset(mode="train", is_interpolate=is_interpolate, target_strategy=target_strategy, mask_sensor=mask_sensor)
train_loader = DataLoader(
dataset, batch_size=batch_size, num_workers=num_workers, shuffle=True
)
dataset_test = AQI36_Dataset(mode="test", is_interpolate=is_interpolate, target_strategy=target_strategy, mask_sensor=mask_sensor)
test_loader = DataLoader(
dataset_test, batch_size=batch_size, num_workers=num_workers, shuffle=False
)
dataset_valid = AQI36_Dataset(mode="valid", val_len=val_len, is_interpolate=is_interpolate, target_strategy=target_strategy, mask_sensor=mask_sensor)
valid_loader = DataLoader(
dataset_valid, batch_size=batch_size, num_workers=num_workers, shuffle=False
)
scaler = torch.from_numpy(dataset.train_std).to(device).float()
mean_scaler = torch.from_numpy(dataset.train_mean).to(device).float()
return train_loader, valid_loader, test_loader, scaler, mean_scaler